On Subjective Uncertainty Quantification And Calibration In Natural Language Generation